The perf subsystem already allows an overflow handler to clear pending_kill
to suppress a fasync signal (although nobody currently does this). Allow an
overflow handler to suppress the other visible side effects of an overflow,
namely event_limit accounting and SIGTRAP generation.

Signed-off-by: Kyle Huey <khuey@kylehuey.com>
---
 kernel/events/core.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index b704d83a28b2..19fddfc27a4a 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -9541,6 +9541,12 @@ static int __perf_event_overflow(struct perf_event *event,
 	 */
 
 	event->pending_kill = POLL_IN;
+
+	READ_ONCE(event->overflow_handler)(event, data, regs);
+
+	if (!event->pending_kill)
+		return ret;
+
 	if (events && atomic_dec_and_test(&event->event_limit)) {
 		ret = 1;
 		event->pending_kill = POLL_HUP;
@@ -9584,9 +9590,7 @@ static int __perf_event_overflow(struct perf_event *event,
 		irq_work_queue(&event->pending_irq);
 	}
 
-	READ_ONCE(event->overflow_handler)(event, data, regs);
-
-	if (*perf_event_fasync(event) && event->pending_kill) {
+	if (*perf_event_fasync(event)) {
 		event->pending_wakeup = 1;
 		irq_work_queue(&event->pending_irq);
 	}
-- 
2.34.1

On Thu, 7 Dec 2023 at 17:35, Kyle Huey <me@kylehuey.com> wrote:
>
> The perf subsystem already allows an overflow handler to clear pending_kill
> to suppress a fasync signal (although nobody currently does this). Allow an
> overflow handler to suppress the other visible side effects of an overflow,
> namely event_limit accounting and SIGTRAP generation.
>
> Signed-off-by: Kyle Huey <khuey@kylehuey.com>
> ---
>  kernel/events/core.c | 10 +++++++---
>  1 file changed, 7 insertions(+), 3 deletions(-)
>
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index b704d83a28b2..19fddfc27a4a 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -9541,6 +9541,12 @@ static int __perf_event_overflow(struct perf_event *event,
>          */
>
>         event->pending_kill = POLL_IN;
> +
> +       READ_ONCE(event->overflow_handler)(event, data, regs);
> +
> +       if (!event->pending_kill)
> +               return ret;

It's not at all intuitive that resetting pending_kill to 0 will
suppress everything else, too. There is no relationship between the
fasync signals and SIGTRAP. pending_kill is for the former and
pending_sigtrap is for the latter. One should not affect the other.

A nicer solution would be to properly undo the various pending_*
states (in the case of pending_sigtrap being set it should be enough
to reset pending_sigtrap to 0, and also decrement
event->ctx->nr_pending).

Although I can see why this solution is simpler. Perhaps with enough
comments it might be clearer.

Preferences?